graph

AI Knowledge Graph (ai/graph)

The graph package is responsible for building a Personal Knowledge Graph (PKG), connecting fragmented Memos into a structured network.

Architecture

The graph consists of Nodes and Edges, supporting multiple types of relationships.

classDiagram
    class KnowledgeGraph {
        +Nodes []GraphNode
        +Edges []GraphEdge
        +Stats GraphStats
    }
    class GraphBuilder {
        -store Store
        -embedding EmbeddingService
        +Build(ctx, userID)
    }
    class GraphNode {
        +ID string
        +Label string
        +Type string
        +Tags []string
        +Cluster int
        +Importance float64
    }
    class GraphEdge {
        +Source string
        +Target string
        +Type string
        +Weight float64
    }

    GraphBuilder ..> KnowledgeGraph : builds
    KnowledgeGraph *-- GraphNode
    KnowledgeGraph *-- GraphEdge

Data Structures

• Nodes: Represent Memo or Tag. Include Importance and Cluster properties.
• Edges:
  • Link: Explicit reference (bidirectional [[WikiLink]]).
  • Tag Co-occurrence: Tag co-occurrence relationship.
  • Semantic: Vector semantic similarity relationship.

Algorithm

1. Building Process (Build)

flowchart TD
    Start[Build Start] --> FetchMemos[Fetch All User Memos]
    FetchMemos --> Nodes[Generate Node List]

    Nodes --> LinkEdges[Build Link Edges (DB Relations)]
    Nodes --> TagEdges[Build Tag Edges (Jaccard)]
    Nodes --> SemEdges[Build Semantic Edges (Vector Search)]

    LinkEdges & TagEdges & SemEdges --> Merge[Merge Edge Sets]

    Merge --> Pagerank[Calculate PageRank Importance]
    Merge --> Community[Calculate Community Clustering]

    Pagerank & Community --> Final[Output KnowledgeGraph]

1. Node Generation: Load all user Memos.
2. Edge Generation:
   • Link Edges: Parse MemoRelation table from database.
   • Tag Edges: Calculate Jaccard similarity between tag sets of any two notes, connect if exceeds threshold.
   • Semantic Edges: For each node, perform Vector Search to find Top-K similar nodes, connect if exceeds threshold.

2. Graph Analysis Algorithms

After graph construction, run graph algorithms to extract insights:

• PageRank: Calculate node importance score. Notes that are referenced more and connected to important nodes get higher scores. Used to control node size in visualization.
• Community Detection: Use simplified Label Propagation algorithm to divide nodes into different clusters. Used to color nodes in visualization, showing knowledge topic distribution.

Filtering

The GraphFilter supports filtering the graph for visualization:

Configuration

Business Flow

1. Frontend requests "Knowledge Graph" view.
2. Backend builds graph in real-time or from cache.
3. Apply filter (ApplyFilter): Filter subgraph by time range, tags, importance threshold.
4. Frontend uses D3.js or other libraries to render Force-Directed Graph.